Multivariate statistical real-time monitoring of an industrial fed-batch process for the production of specialty chemicals
Introduction
Specialty chemicals (e.g. coatings, detergents, resins, adhesives, pigments, additives) are typically obtained by batch or semi-batch processing, where a “recipe” is used to coordinate a sequence of elementary operations (e.g. charge, mix, heat up, react, separate, cool down, discharge), which may be repeated several times during a batch.
Much is reported in the literature on the multivariate statistical monitoring of batch processes in the sense of detection of anomalous or faulty conditions by the analysis of process variables trends (Nomikos and MacGregor, 1994, Kosanovich et al., 1996, Neogi and Schlags, 1998, García-Muñoz et al., 2003, Kourti, 2003, Gunther et al., 2007). Information of this kind is extremely valuable to improve process understanding; however, in most cases it can be used only after the completion of the batch, regardless of the fact it has been obtained in real-time or from a retrospective analysis of the batch itself. For example, if a change in the quality of feeds is diagnosed on-line, the supplier would be interviewed before the next batch is started; if an incipient fouling of a heat exchanger is detected, maintenance would be started at the end of the batch; if the end-point quality of the product is anticipated to be out of specification, the product would not be delivered to the customer, and the reasons for this faulty batch would possibly be analyzed by interrogating the monitoring model. Much less attention has been directed towards the development of monitoring strategies able to provide some kind of information that can be used directly within the same batch from which this information is obtained. In this context, two typical challenges that need to be addressed by a monitoring system in the production of specialty chemicals are the real time estimation of the length of the batch (or the length of any production stage within the batch), and the real time estimation of the instantaneous quality of the product.
There are several specialty productions for which the total batch length is not known a priori, nor is it the length of any processing stage within the batch. Knowing in advance the processing time is useful for several reasons. In fed-batch processes, for example, new fresh material is charged into the process vessels at a convenient time instant. The ability to estimate in real time this instant (which may change from batch to batch) can result in savings both in the number of quality measurements to be processed by the laboratory and in the required total processing time (Marjanovic et al., 2006). On a different perspective, real-time estimation of the total length of the batch can be very useful for production planning, scheduling of equipment use, as well as to coordinate the operating labor resources.
Real time knowledge of the instantaneous quality of the product is extremely important in the manufacturing of several specialties. In fact, if the instantaneous product quality is not found to track a specified trajectory, the processing recipe must be adjusted in real time (possibly several times during a batch), and the batch is kept running until the end-point quality meets the specification. However, in most cases the product quality is not measured in real time, and to contain the laboratory-related expenses (in terms of: need of dedicated personnel, consumption of chemicals, and use analysis equipment) only few product samples are taken during the course of a batch and sent to the lab for analysis. Even so, in a typical industrial scenario where several productions are run in parallel, as many as 15,000 samples may need to be taken and analyzed each year, which can add up to an important fraction of the total product cost. From an operation perspective, due to the scarce number of measurements available during a batch and to the time delay inherent to the analysis, significant drifts on the quality profiles may be experienced before any intervention can be done on the batch being run. The net result is that the recipe adjustments are delayed, the total length of the batch is increased, and the economic performance of the process is further penalized.
In this paper, an industrial case study is presented where the challenges related to the real time estimation of the required processing time and to the instantaneous product quality estimation are addressed using multivariate statistical techniques based on the projection onto latent structures (Wold et al., 2001). Reference is made to a fed-batch process where a resin is produced. It is shown that, by appropriately using existing techniques, information can be extracted from available process measurements that can significantly improve the overall performance of the process. Although the results presented are tailored to a specific case study, we nevertheless believe that the approach we have taken is quite general, and can be useful both to practitioners and to academics for successful design and implementation of data-based monitoring techniques.
Section snippets
The process and related challenges
A high molecular weight polyester resin is produced by catalytic poly-condensation of carboxylic acids and alcohols in an industrial facility where several different specialty chemicals are manufactured. The product quality is determined by the combined values of two indicators, namely the resin acidity number (NA) and the resin viscosity (μ). A schematic of the production process is shown in Fig. 1.
Two distinct production stages are present. Raw materials, catalyst and additives are initially
Projection to latent structures
In order to design the monitoring models, a multivariate statistical technique is used, namely projection to latent structures (PLS), also known as partial least squares regression (Geladi and Kowalski, 1986, Kourti and MacGregor, 1995, Wise and Gallagher, 1996). PLS is a regression technique that allows one to deal with correlated process variable measurements (e.g. temperatures, pressures, flow rates, …) and to relate them to response variables (e.g. quality or time).
Let us first consider a
Stage 1 monitoring: estimation of the stage length
The number of quality measurements within this stage is too small to allow designing a PLS model for the real-time estimation of the product quality. Therefore, to monitor the evolution of the batch, the stage length τ was estimated instead. The possibility of knowing the value of τ in advance would allow the operators to further reduce the number of samples that need taking for analysis, because samples would only be taken starting from the time when the stage is expected to terminate. Thus,
Stage 2 monitoring
During Stage 2 the pre-polymer is transformed into the end product. Interventions on the recipe are carried out by the operators during this stage in response to the quality measurements coming from the lab.
To monitor this production stage two tools were designed: a soft sensor estimating in real time the product quality from process measurements, and a soft sensor estimating in real time the total length of the stage.
Conclusions
In this paper it was shown through an industrial case study how, by considering a blend of engineering judgment and mathematical modeling, multivariate statistical techniques can be exploited to assist the real-time monitoring of product quality and to deliver helpful information for an effective production planning in the semi-batch processing of specialty chemicals.
An evolving PLS modeling approach was exploited for the estimation of the duration of the batch. Namely, it was shown that, by
Acknowledgements
Partial financial support to this work was provided by the University of Padova within Progetto di Ateneo 2005 “Image analysis and advanced modelling techniques for product quality control in the process industry”.
PF, FB and MB would like to thank Sirca S.p.A. for allowing them to publish their industrial data.
References (22)
- et al.
Performance of some variable selection methods when collinearity is present
Chemom. Intell. Lab. Sys.
(2005) - et al.
Online monitoring of multi-phase batch processes using phase-based multivariate statistical process control
Computers Chem. Eng.
(2008) - et al.
Partial least squares regression: a tutorial
Anal. Chim. Acta
(1986) - et al.
Process analysis, monitoring and diagnosis, using multivariate projection methods
Chemom. Intell. Lab. Sys.
(1995) - et al.
Disturbance detection and isolation by dynamic principal component analysis
Chemom. Intell. Lab. Sys.
(1995) - et al.
Real-time monitoring of an industrial batch process
Computers Chem. Eng.
(2006) - et al.
Fault detection properties of global, local and time evolving models for batch process monitoring
J. Process Control
(2005) - et al.
The process chemometrics approach to process monitoring and fault detection
J. Process Control
(1996) - et al.
Modeling and diagnostics of batch processes and analogous kinetics experiments
Chemom. Intell. Lab. Sys.
(1998) - et al.
PLS-regression: a basic tool of chemometrics
Chemom. Intell. Lab. Sys.
(2001)
Cluster analysis for autocorrelated and cyclic chemical process data
Ind. Eng. Chem. Res.
Cited by (25)
Batch-to-Continuous transition in the specialty chemicals Industry: Impact of operational differences on the production of dispersants
2022, Chemical Engineering JournalTroubleshooting an Industrial Batch Process for the Manufacturing of Specialty Chemicals using Data Analytics
2020, Computer Aided Chemical EngineeringCitation Excerpt :Additionally, if the root-cause of the fault causing the product off-specification is not isolated, the product quality issue can manifest again in subsequent batches. Multivariate statistical methods have found many applications in batch process monitoring (Nomikos and MacGregor, 1995; Ündey and Cinar, 2002; Garcia-Muñoz et al., 2003; Faggian et al., 2009). However, most of the reported studies are related to fault detection and isolation in single processing units, whereas plantwide analyses are lacking.
Finding the optimal time resolution for batch-end quality prediction: MRQP – A framework for multi-resolution quality prediction
2018, Chemometrics and Intelligent Laboratory SystemsCitation Excerpt :Among these, methods based on Multiway Partial Least Squares (MPLS) [60], an extension of classical PLS [18], are the most popular models for batch-end quality prediction [49]. They have been applied successfully in a wide range of processes [1,12,14–17,20,35,38,39,44,45,47,56,60,61], yielding estimates of the batch-end quality with accuracy similar to that of off-line laboratory measurements [19]. MPLS models are robust with respect to measurement noise [67], in some cases even surpassing the accuracy of the lab measurements on which they are trained [25,78,79], even though their performance is limited by the uncertainty of the reference measurements in practice [37,54,72].
An error-in-variable projection to latent structure framework for monitoring technical systems with orthogonal signal components
2014, Chemometrics and Intelligent Laboratory SystemsCitation Excerpt :For process monitoring, PLS extracts n sets of latent variables (LVs) that allow constructing a number of non-negative quadratic statistics [6]. Successful applications involving PLS include Refs. [9–13]. For PLS models, the response set typically includes product quality and safety critical variables, while the predictor set contains variables that are expected to directly affect the response set and other recorded variables describing common cause variation.
Probabilistic monitoring of chemical processes using adaptively weighted factor analysis and its application
2014, Chemical Engineering Research and DesignCitation Excerpt :Fault detection and the diagnosis of processes play important roles in the process industry since they are essential in ensuring the safety and improving the productivity of processes. Recently, many multivariate statistical techniques for extracting process information from massive data sets and interpreting this information have been developed in chemical process monitoring (Chiang et al., 2001; Faggian et al., 2009; Qin, 2003; Qin and Yu, 2007; Yoon and MacGregor, 2001), and among these methods, principal component analysis (PCA) and partial least squares (PLS) usually serve as the most fundamental technique and have been successfully applied to monitor a wide range of chemical processes (Alabi et al., 2005; Chiang et al., 2000; Qin, 1998; Kresta et al., 2009; Ku et al., 1995; MacGregor et al., 1994; Nomikos and MacGregor, 1994; Zhang et al., 2011). PCA/PLS-based methods assume that the monitored variables are approximately Gaussian distributed.
Quality assessment of a variance estimator for Partial Least Squares prediction of batch-end quality
2013, Computers and Chemical EngineeringCitation Excerpt :Since PLS-based techniques take information in the outputs into account during model training, they are more suited for regression purposes than PCA-based models. Although PLS has been proven to be a valuable tool in several applications (Facco, Doplicher, Bezzo, & Barolo, 2009; Faggian, Facco, Doplicher, Bezzo, & Barolo, 2009; Gins, Van den Kerkhof, & Van Impe, 2012; Lopes & Menezes, 2003; Marjanovic, Lennox, Sandoz, Smith, & Crofts, 2006), practical implementations in the process industry are limited. A PLS model for batch-end quality prediction describes a linear relation between online process measurements and the final product quality.